The task comes before the hardware
“We need our own AI” sounds decisive. It does not establish what the system should do. Rephrasing a short text differs from searching extensive documents or working on large software projects.
I would therefore put a task list before a hardware list. Only then can deployment choices be compared meaningfully.
A current example of local models
Microsoft's Foundry recap of 15 October 2026 describes Foundry Local 2.1, released on 30 September. It includes support for suitable custom models while stressing runtime and packaging compatibility. [1]
The product documentation identifies on-device processing and offline operation as intended uses. Models must first be available or downloaded. [2] That is not a universal promise that any AI application works without connectivity or other services.
Compare using real tasks first
I would define a few representative examples: revise an original text, answer questions about an approved document or explain a manageable code fragment. Assess result quality, response time and necessary rework.
I would not make the hardest exception the default requirement. A smaller model may be sufficient for frequent tasks. A trial might instead show that the desired work is not practical locally on the available equipment. Either finding is useful.
Running costs extend beyond the purchase
My assessment would include storage, electricity, maintenance time and updates. Model licences and permitted use must fit the task too. No charge per request does not mean free operation overall.
Before buying, examine existing equipment and arrange an appropriate trial. Without that evidence, an expensive hardware recommendation would be premature. This article deliberately offers no universal shopping list.
Local operation is a question about the whole system
I would examine the application's actual data flow. Does it run additional web searches? Are there cloud features, synchronisation or diagnostic uploads? Which files are retained, and who can access them?
A locally running model does not answer these questions for every other component. Combining local and external services may be useful, but the boundary should be deliberate and understandable. Sensitive documents are unsuitable for uncontrolled experimentation.
My conclusion
Local AI is attractive when the task, control over data and operational arrangements fit together. The right solution may be local, cloud-based or mixed. I would decide from tested results, not the label “our own AI”. Vendor statements do not replace assessment of the particular system.
Sources and context
Research date: 24 October 2026. Linked vendor and project information is not an independent product test. Examples and recommendations are identified as such.
[1] Microsoft: September Foundry recap, published 15 October 2026
Read source ↗
[2] Microsoft Learn: Foundry Local and offline operation
Read source ↗